A hybrid GaN HEMT model merging artificial neural networks and ASM-HEMT for parameter precision and scalability

• Main Authors: Lu, Zhongzhiguang, Li, Hanchao, Xie, Hanlin, Zhuang, Yihao, Wang, Wensong, Ng, Geok Ing, Zheng, Yuanjin
• Other Authors: School of Electrical and Electronic Engineering
• Format: Article
• Language: English
• Published: 2025
• Subjects: Engineering; Neural network; Parameter extraction
• Online Access: https://hdl.handle.net/10356/182781
• Institution: Nanyang Technological University

An innovative hybrid physical model for gallium nitride high-electron-mobility transistors (GaN HEMTs) that leverages an artificial neural network (ANN) approach is proposed. This model utilizes ANN to formulate surface potentials, correlating them with the device’s RF and dc behaviors in accordance with the advanced SPICE model for GaN HEMTs (ASM-GaN-HEMTs) theoretical framework. The extraction process is refined through multiobjective particle swarm optimization (MOPSO), enhancing the precision of the extracted parameters. Subsequently, a two hidden-layer ANN architecture is employed to derive the surface potential at the source and drain terminals (ψs and ψd). These surface potentials form the basis of the hybrid model within the advanced-SPICE-model (ASM)HEMT framework, including the trapping and self-heating effects. The validation of the model is conducted using the advanced design system (ADS) simulation platform. The hybrid ANN-based model exhibits scalability and accuracy in comparison to traditional ASM-based physical models. Experimental validations demonstrate a strong concordance between the hybrid model’s predictions and the empirical data across a range of tests, including current–voltage (I–V) characteristics, S-parameters, and load–pull power sweeps. The proposed method significantly improves the accuracy of the S-parameter compared to traditional models, reducing the large signal performance error to within 5%. The results show the robustness of the proposed model and its potential to enhance the predictive modeling capabilities for GaN HEMT devices.